FULL STORY

AI Cracks Millennium Problems, Humbling Expert Forecasts

After LEAP experts gave AI low odds on millennium problems, AI progress defied forecasts, with surveys showing expert timelines rapidly outpaced by reality.

2026-09-09 ~ 2026-09-17 · 3 episodes · 16 posts

Episode 1 · LEAP expert panel puts just 10% odds on AI helping solve a Millennium Prize problem by 2027 (2026-09-09, 6 posts)

The Forecasting Research Institute released the second wave of its LEAP (Longitudinal Expert AI Panel) "AI for Science" report, aggregating forecasts made in August–September 2025 by 277 AI experts, 58 superforecasters, and 1022 members of the public. The headline finding: roughly a 10% probability that AI will independently or substantially help solve a Millennium Prize problem by 2027, signaling expert caution about near-term AI breakthroughs in pure mathematics — a judgment now being tested by actual AI progress.

Confirmed

  • The LEAP panel of AI experts and superforecasters put a 10% probability on AI independently or substantially helping solve a Millennium Prize problem before 2027.
  • Some experts cited a June 2025 statement by the Clay Mathematics Institute president that AI remains far from contributing substantively to problems like the Riemann Hypothesis or P vs NP.
  • The report also included forecasts from a 1022-person public sample as a comparison group.
  • Post materials differ on timing: most recaps describe a "November 2025 panel," while the original report aggregation indicates forecasts were collected in August–September 2025, with November marking the release of results.

Why it matters

  • The low 10% figure highlighted by sharers such as @Afinetheorem and @soumitrashukla9 reflects experts' cautious expectations for short-term AI breakthroughs in frontier pure mathematics.
  • Ethan Mollick (shared by @eldonredwards) counters that AI progress is already outpacing even top forecasters' expectations, underscoring the tension between forecasting methodology and the pace of AI capability gains — a gap worth tracking.

Episode 2 · AI Solves Millennium Problem Far Ahead of Expert Forecasts (2026-09-11, 4 posts)

AI has reportedly cracked a Millennium Prize Problem years ahead of forecasts: a 2025 LEAP report of 277 experts and 58 superforecasters put only a 10% chance of solving one by 2027, with many expecting 2040. Commentators like Eric Topol say the milestone's early arrival shows experts consistently underestimate exponential progress.

Episode 3 · OpenAI's Navier-Stokes result beats expert forecasts by ~30 years (2026-09-16, 6 posts)

The December 2024 ESPAI2024 expert survey by AI Impacts found that typical AI researchers expected a 50% chance of AI solving a Millennium Prize Problem only around 2054, proving a publishable math theorem near 2040, and finding/exploiting security bugs in systems with over 100,000 users in the early 2030s. OpenAI has now announced that its next-generation model produced Navier-Stokes-related results, moving the Millennium-Problem milestone up by roughly 30 years.

Confirmed

  • Expert forecast data from the ESPAI2024 survey published by aiimpacts.org in December 2024 (cited by m3 and m5; source links added by m2 and m4)
  • Forecast details: 2054 for a 50% chance of solving a Millennium Prize Problem; 2040 for a publishable theorem; early 2030s for patching security bugs in large systems
  • OpenAI announced its next model generated results related to Navier-Stokes, a Millennium Prize Problem (m6)

Why it matters

  • benjtodd, kimmonismus and Tolopono all argue this shows AI capability progress far outpacing collective expert expectations
  • Expert forecasts being overturned within under a year suggests estimates of AI progress may be systematically conservative, relevant to long-term planning and policy